Association for Public Policy Analysis & Management

New York City generates approximately 3.2 million tons of waste each year, costing over $1 billion to collect, remove, and send to landfills. The New York City Department of Sanitation (DSNY) provides regularly scheduled curbside recycling and garbage collection for residential buildings, public schools, and city-owned buildings. As part of NYC’s 80x50 sustainability plan, the reduction of landfill waste is an important component of lowering carbon emissions and improving quality-of-life for city residents. Achieving this goal will require new incentives for households to reduce refuse generation through increased recycling and composting activity. Part of the challenge is that DSNY waste data are currently aggregated across 232 “Sections” that divide the City, so there is little information on how much individual buildings generate in terms of refuse and recyclable and compostable materials. This paper develops a machine learning model to predict building-level waste generation at the daily, weekly, and monthly timescales and uses the results to evaluate the impact of several possible incentive and regulatory measures designed to change household behavior.  We use DSNY daily waste collection data over eight (8) years at the route level, which accounts for more than 3,000,000 observations, and integrate these with correlative data, including weather, demographic, socioeconomic, and building/land use characteristics to estimate building- and household-level waste generation rates. In order to predict waste generation per capita across individual truck routes (the highest spatial resolution waste data currently available), we use a Recurrent Neural Network model to develop a spatio-temporal prediction and validate against actual DSNY field-collected waste counts on selected routes conducted specifically for this project. With the results from this model at the building level, we are able to forecast weekly waste generation for each residential property (approx. 800,000) in New York City, and determine expected waste generation based on the type of property, household size, socioeconomic and demographic characteristics and forecasted weather. We apply these results to a peer benchmarking model to design programs that use financial incentives, regulatory levers, and social influence to change household waste generation behavior.